Checkpoints and deletions

Keep these three positions separate:

Position	Owner	Purpose
Execution journal	Browserlayer	Resume completed steps within a run
Source checkpoint	Your sync	Continue fetching the website's dataset
Consumer cursor	Your app	Continue applying published changes

Incremental publication

Inside a sync with a checkpoint schema such as z.object({ next: z.string() }):

const checkpoint = await ctx.getCheckpoint(); // null on the first run
// Fetch the next page through your connector using checkpoint?.next.
await ctx.publish("page", records, {
  checkpoint: { next: nextPageToken },
  deleted: deletedIds,
});

records, nextPageToken and deletedIds come from the upstream page you fetched. Publication commits records, checkpoint and notification intent atomically.

Full snapshots

await ctx.publish("snapshot", records, {
  checkpoint: null,
  scan,
  completeScan: true,
});

Use one unique scan ID for every page in a snapshot. Generate it in a durable step. Only mark the final successful page completeScan: true.

Case	Result
Complete snapshot omits an old record	Publish a deletion
Complete snapshot is empty	Delete all old records and publish an update notification
Partial scan fails	Already published batches remain; missing records are not deleted
Scan continues across runs	Carry its scan ID in the checkpoint

Model changes need a new sync identity or an explicit development dataset reset. Renaming a sync starts a different stream.